The AI ASIC Market, Part 4: The AI Chip Companies Outside the NVIDIA Fight

TLDR

• The independent AI chip market splits between architectural conventionality and commercial maturity — and the highest-profile NVIDIA challengers are not the only viable bets.
• Tenstorrent democratizes AI silicon design by licensing open chip IP — a structurally different bet from selling finished chips, and one targeting markets where hyperscalers don't compete.
• Tensordyne is betting on a chip that replaces multipliers with adders for an 8x power efficiency gain — if first silicon validates the simulation.

In Parts 2 and 3 we mapped two halves of the independent AI chip universe: the design enablers (Broadcom and Marvell) that quietly power hyperscaler ASICs, and the direct NVIDIA challengers (Groq, Cerebras, and Etched) competing for the inference silicon market.

The AI Chip Companies Outside the NVIDIA Fight

Tenstorrent and Tensordyne are independent AI chip companies, but they don't fit cleanly into either of the prior groupings. They're neither design enablers for hyperscalers nor are they fighting head-to-head with NVIDIA for inference deployments. Both are making architectural bets that look different from the rest of the independent landscape, and both are building toward markets that look different too.

That makes them worth covering on their own. Putting them in the same bucket as the Part 3 cohort produces wrong conclusions about how to evaluate them — Tenstorrent and Tensordyne operate on different strategic logic, and the gating variables that determine their success are different too.

Tenstorrent: The Open RISC-V Alternative

Tenstorrent's strategy is the most differentiated among the independent AI chip companies. The company isn't trying to build a better closed GPU. It's trying to build a better open ecosystem around RISC-V and efficient AI acceleration. The thesis is to democratize AI silicon design, which is a different business than selling finished chips.

Led by Jim Keller — architect behind Apple's A-series CPUs, AMD's Zen microarchitecture, Tesla's FSD chip, and multiple Intel projects — Tenstorrent has built a product stack that reflects this positioning. Tensix cores handle AI processing for training and inference. The Blackhole and Wormhole processors are current-generation AI accelerators. But the strategically important part is the Ascalon RISC-V CPU plus Tensix AI IP licensing. Tenstorrent licenses the RTL source code — the actual circuit description — rather than black-box licensing like ARM. This is an order of magnitude more flexibility for customers who want to design their own chips without building a full silicon design team from scratch. Galaxy systems are claimed 3x more efficient and 33% less expensive than NVIDIA DGX. Quasar is an upcoming Samsung-manufactured product that demonstrates manufacturing diversification beyond TSMC.

Jim Keller's framing on memory cost is the clearest articulation of Tenstorrent's positioning. He has publicly argued that competing with NVIDIA on HBM-based architectures is structurally difficult — NVIDIA buys the most HBM, has the largest cost advantage, and has the deepest supply relationships. Tenstorrent designs around lower-cost memory deliberately, targeting markets on cost and efficiency rather than peak FLOPS. It's a strategy that wouldn't work if the goal were to win the frontier training market. But Tenstorrent isn't targeting the frontier training market.

The customer base reflects the diversification strategy. Signed customer contracts total approximately $150 million across LG Electronics (Ascalon CPU plus video processing), Hyundai/Kia (automotive AI co-processors), and Samsung (manufacturing plus automotive). This is durable revenue spread across automotive, consumer electronics, and enterprise edge — not the high-profile hyperscaler deals that dominate Groq and Cerebras headlines, but arguably more diverse and resilient. Tenstorrent's valuation reached $3.2 billion in a December 2024 Series D round.

The strategic positioning is genuinely different from every other independent AI chip company in this series. Tenstorrent is carving out markets hyperscalers don't serve — edge AI, automotive, mid-tier enterprise, and the open IP licensing market itself. This is largely non-competitive with hyperscaler in-house silicon, which is optimized for data center workloads. The RISC-V IP strategy also positions Tenstorrent as a potential foundational supplier to companies that want to build their own chips without going through ARM or Broadcom. That's a structurally different opportunity than competing with NVIDIA on data center inference, and it may be a more defensible long-term position even if it produces lower headline valuations.

Tensordyne: The Logarithmic Outlier

Tensordyne makes the most technically radical claim in the independent AI chip market. By computing AI inference in the logarithmic domain rather than conventional floating-point or integer arithmetic, the company says it can reduce chip area, power consumption, and capital cost by factors that dwarf competing approaches.

If the claims hold in production silicon, Tensordyne represents a step-function change in AI inference economics. The word if is doing a lot of work in that sentence.

The company was founded in 2017 as Recogni, focused on automotive edge inference. The September 2025 rebrand to Tensordyne signaled a full pivot from automotive edge to data center inference. The company is still pre-revenue at scale.

The architectural thesis, called the Pareto number system, rests on a specific mathematical observation. Matrix multiplication — the dominant neural network operation — becomes simple addition in the logarithmic domain. Adder circuits are significantly smaller and less power-hungry than multipliers. This frees up die area for more SRAM cache (Tensordyne claims 6x more than GPU equivalents), which improves throughput. The key challenge is that addition itself becomes more complex in the log domain; Tensordyne addresses this with a proprietary approximation method.

The product specifications are promising but entirely simulation-based. Two numbers do most of the work: 1/8th the power per token versus NVIDIA GB200 NVL72 racks, which would change the economics of every power-constrained data center site, and air-cooled operation, which would eliminate the liquid cooling infrastructure GPU racks require. Tensordyne also claims 1/3rd the capex per token and reports less than 1% accuracy error on most models. None of this has been validated in silicon yet.

Tensordyne achieved IDCA G2 certification in September 2025 — an independent validation of enterprise AI platform operational readiness — and describes hyperscalers and neo-cloud companies as "lined up" for beta testing. No named customers are publicly confirmed. CEO Marc Bolitho described chip tape-out as "imminent" in late 2025, with hardware launch targeting mid-2026.

The risk profile is the sharpest of any company in this series. All performance and efficiency figures are simulation-based; first silicon must be validated before any of the claims can be tested. Logarithmic number system architectures have been researched for decades without mainstream adoption, and Tensordyne's approximation solution to the addition-in-log-domain problem is proprietary and unproven at production scale. The fact that this architecture hasn't shipped in a data center chip previously is evidence, not just skepticism. Tensordyne is also significantly less capitalized than peers, with most funding under the prior Recogni brand before the data center pivot, and no named anchor customers are publicly committed. Beta testing interest is not the same as purchase orders.

That said: if Tensordyne's claimed 8x power efficiency holds in production silicon, the value proposition for data centers facing energy constraints would be significant. The "if" is enormous. But an 8x power efficiency advantage, if real, changes the economics of AI inference deployment in ways that matter — especially for edge and mid-tier enterprise deployments where liquid cooling infrastructure is prohibitively expensive.

Tenstorrent vs. Tensordyne: How the Two Architectural Outliers Compare

Tenstorrent and Tensordyne sit in the same broad category — independent AI chip companies operating on different strategic logic from the head-to-head NVIDIA challengers — but the distinctions between them are sharper than the shared category suggests.

The two companies are betting on opposite ends of the silicon-conventionality spectrum. Tenstorrent's RISC-V foundation is the most conventional silicon choice in the independent AI chip landscape — what makes Tenstorrent radical is the open IP business model, not the chips themselves. Tensordyne is the inverse: its business model (sell finished accelerators to enterprise data centers) is conventional, but the underlying logarithmic number system is the most architecturally radical bet of any company covered in this series.

The commercial maturity gap is just as sharp. Tenstorrent has shipping silicon, $150 million in signed contracts, and a $3.2 billion valuation built on real revenue. Tensordyne has simulation benchmarks, beta testing interest, and a tape-out described as "imminent." Both are credible bets, but they're at completely different points on the validation curve.

Where the Architectural Outliers Sit in the Independent AI Chip Market

Plotting Tenstorrent and Tensordyne alongside the Part 3 cohort — Groq, Cerebras, and Etched — shows the variety of approaches in the independent AI chip category:

Tenstorrent sits in the conventional/validated quadrant alone — every other independent AI chip company in this series has placed a more architecturally radical bet, and Tenstorrent is doing well by deliberately not competing on that axis. Tensordyne sits in the radical/speculative quadrant alone, with the most extreme architectural concentration risk and the least commercial validation in the entire group.

The cluster in the radical/validated quadrant — Groq, Cerebras, and Etched — is where most coverage focuses. That's where the $20B acquisitions and $35B IPO targets live. But the grid shows that two of the five independent AI chip companies don't sit in that cluster, and the strategic logic of those two outliers is worth understanding on its own terms. Not every viable AI chip company is a head-to-head NVIDIA challenger.

What's Next: The Synthesis

Parts 2, 3, and 4 have mapped the independent AI chip universe — design enablers, head-to-head NVIDIA challengers, and architectural outliers. One last thread worth pulling for AI/HPC pivot operators: if Tensordyne's claimed 8x power-per-token advantage holds in production silicon, the deployment math shifts toward sites with constrained power and air cooling — exactly the profile of most Bitcoin mining infrastructure. The "if" is enormous, but every architectural advance that lowers AI inference power requirements expands the set of sites where mining infrastructure can support AI workloads without major retrofits.

Part 5 closes the series with the synthesis: the broader investment landscape, NVIDIA's strategic response to the independent challenger group, and the structural conclusions that emerge when all five parts are read together.